En DashHotdogBenchmark
Source on GitHub: fork it, clone it

Every week, asked again.

Newest first. An edition is never revised after it is published. The cost is the sum of every call's estimate from the dated price table, so the bill for asking is a number, not a guess.

Published editions
EditionPublishedQuestionsModelsEst. costNotes
Week 42, 2026October 12, 2026612$0.46
Week 40, 2026September 28, 2026612$0.45
Week 39, 2026September 21, 2026612$0.45
Week 38, 2026September 14, 2026612$0.45
Week 37, 2026September 7, 2026612$0.45
Week 36, 2026September 3, 2026612$0.47

Each question, week by week

How every model has answered across editions: verdict share, latency, tokens and framing sensitivity over time.

What each edition found

The control tally per question: the question asked plainly, no system prompt. The framed arms and every model's answer are in the edition itself.

Week 40, 2026

Published September 28, 2026. 12 models, 3 framings, 6 questions.

Week 38, 2026

Published September 14, 2026. 12 models, 3 framings, 6 questions.

Week 36, 2026

Published September 3, 2026. 12 models, 3 framings, 6 questions.